Anthropic Releases MHS: Claude Controls Robotic Arms and Quantum Computers

新智元 · wechat · 2026-08-28

Anthropic has released a research preview of the Model Hardware Standard (MHS), enabling AI to directly control physical world hardware, dubbed the 'hardware version of MCP.'

Core Functionality:

MHS allows models like Claude to control robotic arms, microscopes, and laser systems via a standard interface.
AI can master unfamiliar machines by reading simple hardware descriptions (e.g., speed, temperature limits), eliminating the need for custom integration code.

Case Studies:

Quantum Calibration: At QuEra, Claude used MHS to tune laser parameters, achieving 99.3% stability (vs. 58% by humans) and reducing calibration time from 5-10 minutes to 6 seconds.
Drug Discovery: At Genentech, Claude autonomously optimized pipette settings and fixed bugs. While it struggled with physical intuition issues like bubbling liquids, it maintained a closed-loop workflow.
Brain Imaging: At HHMI Janelia, Claude coordinated multiple devices, compressing a weeks-long experiment into a single day.
Zero-Shot Arm Control: Hugging Face researchers found that untrained Claude Code directly controlled a robotic arm to sort colored blocks.

Limitations:

Lack of Physical Intuition: AI may loop inefficiently when facing physical obstructions or liquid foaming.
Device Coverage: Only supports hardware with programmable interfaces; legacy equipment is often excluded.
Cost & Regulation: Continuous monitoring incurs high compute costs, and incorrect safety boundary descriptions pose liability risks.

Original post →

Anthropic Releases MHS: Claude Controls Robotic Arms and Quantum Computers

More from Embodied